When
Where
Speaker: Weimin Zhou, Radiology and Imaging Sciences and Optical Sciences, University of Arizona
Title: Computational and Learning-Based Methods for Ideal Observers and Task-Based Optimization of Image Quality
Abstract: Modern imaging systems produce images through complex imaging chains involving numerous design parameters. To evaluate and optimize imaging systems for a specific task, such as tumor detection in medical imaging, task-based measures of image quality are needed that quantify the performance of an observer on that task. The Bayesian Ideal Observer (IO) provides an upper bound on observer performance, which serves as a principled figure of merit for task-based image quality assessment. However, the IO employs a likelihood ratio for decision making, which is often analytically intractable and can be computationally challenging to evaluate for high-dimensional image data. In this talk, I will introduce computational and learning-based methods for addressing these challenges. I will first describe how deep generative models can be used to learn stochastic object models that characterize object variability underlying the distribution of imaging data, and then show how these models can be leveraged with Markov-chain Monte Carlo methods to approximate the IO. I will also describe a conjugate gradient-based channelization method for task-based image dimensionality reduction that facilitates IO computation. Finally, I will introduce an image-adaptive deep learning strategy that employs task-based channels to guide the optimization of neural networks for image denoising. Together, these computational and learning-based approaches enable task-based assessment and optimization of imaging systems.